Validation of freely-available pitch detection algorithms across various noise levels in assessing speech captured by smartphone in Parkinson’s disease

Abstract Measuring the fundamental frequency of the vocal folds F0 is recognized as an important parameter in the assessment of speech impairments in Parkinson`s disease (PD). Although a number of F0 trackers currently exist, their performance in smartphone-based evaluation and robustness against background noise have never been tested. Monologues from 30 newly-diagnosed, untreated PD patients and 30 matched healthy control participants were collected. Additive non-stationary urban and household noise at different SNR levels was added to the recordings, which were subsequently assessed by 10 freely-available and widely-used pitch-tracking algorithms. According to the comparison of all investigated pitch detectors, sawtooth inspired pitch estimator (SWIPE) was the most robust and accurate method in estimating mean F0 and its standard deviation. However, at a low 6 dB SNR level, a combination of more algorithms may be needed to achieve the desired precision. Monopitch, calculated as F0 standard deviation and estimated by SWIPE, proved to be robust in distinguishing between the PD and healthy control groups (p